Fuente: Wipo "digitalization" A predicted digital image is generated (1602) from a semantic segmentation of a digital image using a Neural Network (NN), then a realism prediction is generated (1608) utilizing a semantic discriminator (see [0031], fig. 5 & [0096]) from the predicted digital image. The NN’s parameters are modified (1610) based on the realism prediction (e.g. using adversarial losses, e.g. [0090], [0160]). Generating a semantic image embedding from the predicted image and the semantic segmentation and using the embedding to generate a realism prediction is claimed. Generating a realism prediction from image-level and object-level semantic discriminators is also claimed. First and second encoders used to generate an image embedding from the predicted image, and a semantic image embedding from the predicted image and the semantic segmentation respectively is also claimed. Generating the realism prediction may comprise using the semantic discriminator and a generative adversarial discriminator to generate first and second realism scores respectively. A realistic inpainted digital image is constructed which also conforms to a semantic segmentation. A user interface to facilitate inpainting interactions is also disclosed (Fig. 6A-D & 7A-D). Iteratively updating an inpainted digital image based on changes to a segmentation map is also disclosed (210 Fig. 2).